ANOMALY DETECTION AND REMOVAL IN RADAR IMAGES (ANDRE) FINAL PROJECT REPORT. Roskien poisto tutkakuvista (ROPO) projektin loppuraportti

Transcription

1 ANOMALY DETECTION AND REMOVAL IN RADAR IMAGES (ANDRE) FINAL PROJECT REPORT Roskien poisto tutkakuvista (ROPO) projektin loppuraportti Markus Peura, Finnish Meteorological Institute, Development Branch P.O. Box 503, Helsinki, Finland Project manager: Jarmo Koistinen October 18, 2002 ABSTRACT The goal of this project was to detect and remove the anomalies appearing in weather radar data. The main targets were sea clutter (ships, waves), interfering external emitters (telecommunication, other radars) as well as the sun, birds, insects, and aircraft. Methodologically, the strategy was to use computer-vision related techniques as well as simpler methods applying auxiliary meteorogical data sources. A central principle was to keep detection and removal as distinct processes. This design is meaningful when 1) generating products from the anomalies, when 2) generating radar image products for which user may set the degree of removal, and 3) for data fusion involving dynamic source weighting (by quality or confidence). The related software has been in operational use at the FMI since May The scientific novelties are explained in further detail in the related conference publication (Peura 2002a). 1 INTRODUCTION Weather radar images are frequently contaminated by non-meteorological echoes as well as echoes caused by anomalous propagation. Especially in spring, the temperature differences between sea and the lowest atmosphere causes radar beams to bend downwards, yielding strong echoes from sea waves and ships as well as other anomalies like second trip echoes and external emitters. Some anomalies are shown in Fig. 1. lowest sweep (0.5 ), original B-scan lowest sweep PPI, map composite Fig. 1. An example of radar data containing several types of anomalies FMI Korppoo radar at 04:30 EET on 9th July, The only actual precipitation is in the North / North-East. The anomalies are speckle noise, sun (the continuous line segment pointing to 40 ), emitter (170 & 280 ), insects (near the radar), ships (sharp small specks; sidelobes), and sea clutter (large speck between 180 and 270 ). The detectors presented in this paper will be demonstrated using this data.

2 Automated detection of the various anomalies has proven to be a difficult task. However, even if the performance of an automated detection scheme remains inferior to the quality of a meteorologist s analysis, there are still advantages like speed, consistency, and objectivity. One should also keep in mind that the requirement for the purity of data differs among applications. For example, operational generation of warnings often requires conservative anomaly removal, if any, whereas in computing motion vectors pure radar data samples are preferred even at the cost of spatial coverage. Practically, our computer vision based approach means that the targets are treated above all by their visual appearance. We group and discuss targets primarily under the detection algorithms expected to work on them. For example, small specks orginate from various sources, but a filter that removes them all in a single pass will do. However, in many cases, similar appearance tends to refer to similar origins. The detector set developed at the FMI is shown in Table 1. Table 1. Detectors and their primary targets. detector target BIOMET birds and insects near the radar SPECK noise; distinct specks EMITTER line segments SUN long line segments SHIP ships (and aircraft) VERT_GRAD* sea waves and ducting effects METEOSAT* suspiciously warm data DOPPLER* non-continuous doppler data * not operational (yet) In the filtering process, we treat detection and removal as separate processes. This principle is motivated by two aspects of efficient end-product computation. First, some anomalies are of interest of certain customer groups, such as birds for aviation. Second, products have varying requirements for the purity of data. Hence, the results of detection are presented as images of continuous-valued probabilities, enabling several products to be generated from a single original by repeated thresholding. This scheme is illustrated in Fig. 2. This philosophy is compatible with the data management and control design proposed for the new weather prediction system at the FMI (Kukkonen and Kallio 2002). Fig. 2. Overall scheme. Each detector processes input data and produces an independent detection map, probability of anomaly. 2

3 2 METHODS The principal image analysis techniques applied in this project are listed below. Selected techniques illustrated in Fig. 3 and Fig Mask operations. A variety of digital image processing tasks can be performed by means of masks (also called windows or templates). The mask techniques applied in this project are related to the common median filtering. Such detectors are efficient in cancelling sharp details such as outliers (say, ships in radar imagery) and speckle noise. (Haralick et al. 1987). 2. Morphological operations. In computer vision, morphology refers to operations causing expansion or retraction of shapes, typically not as a result of geometrical scaling but by so called erosion and dilation. In nature, repeated freezing and melting produce analogous results. In this project, morphology has been used in collecting evidence of anomalies by connecting adjacent segments. (Haralick et al. 1987, Sonka et al. 1993). sample image f(x, y) morphol. operation (horizontal closing) segment operation (speck area) segment operation (horizontal run lengths) Fig. 3. Basic image analysis operators. 3. Segment operations. Many graphics programs apply the flood-fill algorithm in painting objects automatically. This technique can be also used in spreading information in an image. One may for example scan all each segment in the image, compute its area (closed curve integral), and then spread the result within the segment. This way, further processing, say filtering too small specks in radar images, can be performed simply by thresholding and masking. If the objects of interest are elongated, it is reasonable to distinguish them by their runlengths, that is, by the segment lengths calculated in either horizontal or vertical direction. (Sonka et al. 1993, Gonzales and Woods 1992, Peura et al. 1998) 4. Fuzzification. Targets appearing in meteorological radar data different modes of precipitation, bright bands, sky conditions and anomalies cannot be separated by applying strict dbz thresholds. Hence, in detecting and classifying these targets we suggest producing smooth curves of probability (or certainty, confidence, quality, or Bayesian belief) instead of absolute if-then results. This approach also helps in keeping results independent from scaling and measuring units. We propose using soft peak and threshold functions for communicating meteorologist s expertise (Fig. 4). For motivation, consider translating the following sentence to a mathematical form: if the size of the image segment is around 8 pixels, or at least between 4 and 12 pixels, and its maximal intensity is over 40dBZ, then it is probably a ship δ (x, a) = a2 a 2 +x 2 fuzzy peak -a 0 a σ (x, a) = x a+ x fuzzy threshold -a 0 a Fig. 4. Suggestions of simple fuzzy functions. In both functions, a is a steepness parameter referring to a half-width. The algorithmic design of the detectors is explained in in the conference publication (Peura 2002a). 3

4 3 DETECTION RESULTS This section addresses the current status of the detectors listed in Table 1. Efficiency of the detectors is illustrated in Fig BIOMET birds and insects. In the Finnish radar network, birds and insects ( biometeors ) appear regularly from spring to autumn. Biometeors form low-intensity speckled patterns near the radar. The applied detector is fairly simple - it directly fuzzifies these properties (low intensity and proximity). The results shown in Fig. 8 were obtained using threshold altitude (i.e. 50%-confidence) of 2500m and and threshold intensity of -5dBz. The detector is cabable of indicating the locations of biometeors relatively well but it cannot currently distinguish between small convective cells and dense flocks of birds. Although one could consider modelling the behaviour of biometeors (for birds, see (Koistinen 2000)), designing a fully automatic detector is an ambitious project. 2. SPECK speckle noise and distinct specks. Applying the segment methods explained in the previous section detection of distinct specks is straighforward. One simply sets a threshold for the segment size (in pixels, i.e. bins) corrensponding to the 50% probability of anomaly, then computes the segment sizes. It should be emphasized that segment-size based filtering is more accurate and faster than median filtering usually applied for similar tasks. 3. EMITTER and SUN line segments. In the Finnish radar network, the most frequent type of interfering radiation appears as straight line segments cause by other emitters. However, due to geometry, distant precipitation may appear more or less similarly. The challenge is to distinguish between these two. The detector is designed on the following assumptions: 1) emitter anomaly appears as horizontal segments of length at least a pixels, 2) such segments have maximal vertical width of w pixels (=degrees), and 3) occurrence of short, beam-wise segments implies increased probability of larger emitter anomalies in that direction. The SUN detector is similar to EMITTER with the exception that the expected region in the image the direction of the sun is precomputed. Fig. 5. Intermediate processing steps in detecting emitter segments with EMITTER (compare this with the b-scan image of Fig. 1). From left to right: 1) segments of vertical width 1 and the respective 2) thresholded averages of each beam, and 3) horizontal run lengths from which vertical run lengths have been subtracted. Fig. 6. Intermediate processing steps of SHIP. Left to right: 1) suspicious pixels extracted by simple high-boost filtering, 2) the previous result with horizontal segments pruned, and 3) artificially generated sidelobes to be matched with the actual ones in the input image. 4. SHIP marine (and airborne) vessels. As ships and aeroplanes are efficient reflectors of electromagnetic radiation they appear as sharp spots in weather radar images. The challenge is to distinguish these spots from 4

5 small convective cells; otherwise they risk being interpreted and forecasted as local showers or hail. The SHIP detector is currently the computationally heaviest anomaly detector applied at the FMI. The processing time for one radar volume remains still around a few seconds on the current Pentiums. 5. VERT_GRAD sea clutter detector. Typically, sea clutter interferes the lowest sweep(s) and disappears relatively sharply on the upper sweeps. On the contrary, precipitation echoes have smoother vertical gradients. Hence, our strategy is to consider the vertical gradient which is, however, fuzzily weighted by its altitude; the weight is inversely proportional to altitude. In addition, we should somehow cancel already-passed positive gradients, hence we always compute the gradient from the mimimum dbz value obtained that far. We used 50%-limits 3500m for altitude and -10dBZ/1000m for vertical gradient, and obtained the results shown in Fig METEOSAT 2nd trip, ducting and sea clutter. Precipitation implies ice particles hence too warm spots can be detected anomalous. Because of the low resolution of geostationary satellite data in the latitudes of Finland, the temperature of the smallest specks in radar images cannot be verified. Hence, we mask them out as a preprocessing stage. The results in Fig. 8 were obtained with a 50%-confidence threshold -5 C; respectively 0 C for a 75%-confidence. 7. DOPPLER birds and other indepedently moving targets. As opposite to precipitation, some anomalies have inconsistent doppler velocity fields. Especially insects, dense flocks of birds as well as some sea anomalies appear as discontinuities in doppler data. The results in Fig. 8 were obtained by detecting discontinuities in a 3 3pix window. Finally, in Fig. 7 we show the original image from which all the detectors listed in this paper have been applied. Variating the filtering threshold illustrates how a user can achieve a desired POD/FAR-type compromise. Fig. 7. Collective application of all the listed detectors with 40%, 50%, 75% and 90% filtering thresholds. 5

6 detector response (b-scan) extracted anomalies filtered image BIOMET SPECK EMITTER SHIP VERT_GRAD METEOSAT DOPPLER Fig. 8. Detection results for the case shown in Fig. 1. 6

7 4 CURRENT OPERATIONAL STATUS The current operational mode of the filter set applies brute force removal policy for anomalies: every pixel for which the confidence of anomaly is over 50% is simply erased from the image. One may think the resulting product as the best available radar data. One may however keep in mind that the threshold of compromise (erase anomalies vs. erase rain) varies in applications and users. 5 POSSIBLE FURTHER DEVELOPMENT In image restoration, one could also use more intelligent anomaly removal. Currently, removal of some anomalies found within precipitation may leave black spots in precipitation. One could for example spread the data from neighboring pixels to fill the gaps caused by erasing anomalous pixels. There are three detectors which would need further development before they can be taken to operational use: VERT_GRAD, METEOSAT and DOPPLER. The last detector (for birds, potentially) is technically the simplest and hence easiest to introduce. On the other hand, the current scheme generates the following intermediate products which are however not used as input for any end product. Given a radar volume as input, the currently unused intermediate products are: anomaly map (the most probable anomaly for each data point), confidence map for each anomaly, and total confidence map. These intermediate products could be used for example for generating: adaptive radar composite images weighted by data quality, nowcasting based on images weighted by data quality, radar images with probable anomalies marked radar images with anomalies made transparent or gray (smoothly based on their confidence), and detection products, for example bird maps. None of the detectors uses a sequence of images as input. No previous results of anomaly detection are utilized either. In future, these resources could be used as well. 6 DISCUSSION We presented a set of anomaly detectors for weather radar images. The detectors share a common basis in a set of image analysis techniques, but otherwise the detectors vary in performance and computational complexity. The performance of most detectors, especially for EMITTER, SUN and SHIP is satisfactory. Also the simple BIOMET works well in detecting the existence of birds and insects. The currently missing detector for strong bird echoes might be realized in the future based on the currently available biometeor support field. SPECK provides an exact means for handling invidual specks a property not achieved when using standard median filtering. Some detectors, like the METEOSAT, seem promising, but however do not seem to solve their proposed detection tasks autonomously but hopefully with a help of some additional information available in the future. 7 REFERENCES Gonzales, R. C. and Woods, R. E. (1992). Digital Image Processing, Addison-Wesley. Haralick, R. M., Sternberg, S. R. and Zhuang, X. (1987). Image analysis using mathematical morphology, IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-9(4): Koistinen, J. (2000). Bird migration patterns on weather radars, 1st European Conference on Radar Meteorology, Bologna, Italy, Vol. 25 of Physics and Chemistry of the Earth (Part B: Hydrology, Oceans and Athmosphere), European Geophysical Society, Pergamon Press, pp

8 Kukkonen, P. and Kallio, A. (2002). Pro seija - sääennustusprosessin uudistus. Private communication, (FMI Intranet). Peura, M. (2002a). Computer vision methods for anomaly removal, Second European Conference on Radar Meteorology (ERAD02), Copernicus Gesellschaft. Peura, M. (2002b). An interface for monitoring radar images. operator/ropowww/browser.php (FMI Intranet). Peura, M. (2002c). Ropo-ohjelmisto. peura/ropo/ (FMI Intranet). Peura, M., Koistinen, J. and King, R. (1998). Visual modelling of radar images, COST-75: Advanced weather radar systems international seminar, European Commission, pp Sonka, M., Hlavac, V. and Boyle, R. (1993). Image Processing, Analysis and Computer Vision, Chapman & Hall Computing. 8